DEVELOPMENT AND VALIDATION OF A MACHINE LEARNING TOOL FOR EARLY DIAGNOSIS OF INFLAMMATORY BOWEL DISEASE IN THE PRIMARY CARE SETTING: A POPULATION BASED STUDY. (22nd January 2022)

• Record Type: Journal Article
• Authors: Ber, Tahel Ilan; Tov, Amir Ben; Gazit, Sivan; Steinberg-Koch, Shlomit; Getz, Benny; Jenudi, Yonatan; Underberger, Dan; Ramni, Or; Ben-Horin, Shomron
• Is Part Of: Inflammatory bowel diseases. Volume 28(2022)Supplement 1
• Page Start: S20
• Page End: S21
• Publication Date: 2022-01-22
• Subjects: Inflammatory bowel diseases -- Periodicals; Colitis, Ulcerative -- Periodicals; Crohn Disease -- Periodicals; Inflammatory Bowel Diseases -- Periodicals; 616.344
• DOI: 10.1093/ibd/izac015.030
• Languages: English
• ISSNs: 1078-0998

Abstract

Abstract: BACKGROUND: Although 3, 000, 000 Americans are afflicted with Inflammatory Bowel Disease (IBD) encompassing Crohn's Disease (CD) and Ulcerative Colitis (UC) 1, timely diagnosis is a challenge due to non-specific and overlapping symptoms. 2 More than 20% of patients may be initially misdiagnosed 3 causing delayed diagnosis and treatment, potentially leading to increased risk of complications 2 and irreversible mucosal damage. 4 Artificial intelligence (AI) models can alert physicians to patients who would otherwise be misdiagnosed, potentially improving patient outcomes and reducing costs. We aimed to develop PredictAI, a proprietary AI Gradient Boosted Decision Tree based machine learning algorithm and test its accuracy in identifying undiagnosed CD and UC in the primary care setting. METHODS: This was a retrospective study of 2, 471, 267 patients' electronic medical records (EMR) from Maccabi Healthcare Services in Israel. Sufficient data was available between the years 2010-2020, of which 2 consecutive years (2015-2016) were pre-assigned to the test set. Inclusion criteria were: (i) CD or UC ICD code as defined by Maccabi's Registry, 5 (ii) no other autoimmune disease diagnosis, (iii) at least 4 years of data antedating first suspicion by primary care physician (PCP) of IBD were available. First suspicion was defined as any diagnostic test, procedure, or referral to a specialist, indicating suspicion of IBD. Here we included adult data only. RESULTS: Of 2, 471, 267 patients, 1, 214 had a first-time diagnosis of IBD and available antedating data in the years 2015-2016. Of these, PredictAI identified 126, 120, 104, 93 patients 1, 2, 3 and 4 years prior to initial PCP suspicion, respectively. For CD, it predicted 83/229 (36%) of patients 1 year prior to PCP initial suspicion of disease, 70/211 (33%) patients 2 years prior, 60/179 (33%) patients 3 years prior and 56/148 (38%) patients 4 years prior. Discriminatory accuracy area under the curve (AUC) was 76%, 75%, 75% and 78%, 1, 2, 3 and 4 years before initial PCP suspicion, respectively (Figure 1). Corresponding early-identification ratio and discriminatory accuracy for UC was 42/238 (17%, AUC=71%), 50/225 (22%, AUC=70%), 44/187 (23%, AUC=72%) 37/161 (23%, AUC=74%), for 1, 2, 3 and 4 years before initial PCP suspicion of UC, respectively (Figure 2). Specificity for CD and UC each was above 90%. CONCLUSIONS: PredictAI accurately identified CD and UC diagnosis in 17-38% of patients presenting to primary care up to 4 years prior to PCP's initial suspicion, potentially reducing time to diagnosis.